Terminal-Bench 2.0

Coding

Measures how well AI agents complete real tasks in a terminal: software engineering, ML, security and data science, each run in a containerized sandbox. Higher is better.

Official source

Terminal-Bench 2.0 contains 89 verified tasks that each run in a containerized terminal environment via the Harbor framework. An agent gets a task instruction, works inside the sandbox, and the run is scored pass or fail by task-specific checks, so the headline number is the percentage of tasks resolved, published with a confidence interval. Every entry is an agent harness plus model pair, not a bare model: the same model scores differently depending on the scaffold wrapping it (Codex CLI, Claude Code, Droid, the reference Terminus 2 agent and others), with harness quality worth several points. We list each model's strongest harness rows with the harness named on the row; submissions are verified by the Terminal-Bench team.